Forecasting macro-economic variables using relevant predictors

This paper expands on the factor model based forecasts by allowing for more exibility in the factor model. First a non-linear relationship between the predictors and the factors is allowed by expanding the dataset with the squared predictors. Second a subset of relevant predictors is selected based on their association with the series to be forecast. Hard thresholding, LARS and adaptive lasso techniques are used to select the relevant predictors. Factors are then estimated by applying prin- cipal components to these relevant predictors. These relevant predictors change for dierent dependent variables, forecast horizons and over time. Improvements over the normal factor model forecasts are found for all forecast horizons and variables.